What Offline Training Can and Cannot Teach

You can study every published map of a city and still be surprised by a water main break. Offline training has that limit. It can show a network a huge archive of how people drove in the past. It cannot include an event that has not been captured yet.

What offline training is good at: repeating patterns, smoothing ordinary merges, reading common signals, copying competent following distance in familiar weather. What it cannot promise: correct behavior in a brand-new layout, a unique vehicle, or a combination of rain, glare, and a missing sign that never co-occurred in the set.

Offline training has that limit.

Teams try to fake the missing future with simulation and with aggressive mining of near-miss clips. Both help. Neither is a time machine. The honest statement is narrower. Offline training teaches the distribution you already have. Driving in the world keeps generating a slightly different distribution. Closing that gap is the loop later posts call fleet learning.